Research map / Image generation / Fast diffusion sampling
Energy-Clibrated VAE with Test Time Free Lunch
Yihong Luo, Siya Qiu, Xingjian Tao, Yujun Cai, Jing Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add energy-based calibration to VAE training (ablation) | VAE · CIFAR-10 | FID: 104.68 -> 5.20 | Improved |
| add MCMC sampling at test time (ablation) | EC-VAE · CIFAR-10 | FID: 5.20 -> 5.63 | No clear effect |
| add MCMC sampling at test time (ablation) | EC-Flow · CIFAR-10 | FID: 31.12 -> 22.74 | Improved |
| calibrate the normalizing flow with energy-based calibration (ablation) | Glow · CIFAR-10 | FID: 91.33 -> 31.12 | Improved |
| use primal-dual optimization instead of plain SGD with fixed weight (ablation) | EC-VAE · CIFAR-10 | FID: 5.72 -> 5.20 | Improved |
| use adversarial WGAN calibration instead of energy-based calibration (ablation) | VAE · CIFAR-10 | FID: 5.20 -> 33.78 | Hurt |
| initialize EBM with pretrained VAE samples (ablation) | EBM · CIFAR-10 | FID: 45.52 -> 44.63 | No clear effect |
| add flow prior to EC-VAE (ablation) | EC-VAE · CIFAR-10 | FID: 5.20 -> 4.85 | Improved |
| add LPIPS to decoder likelihood (ablation) | EC-VAE · CIFAR-10 | FID: 5.20 -> 4.77 | Improved |
| add calibrated posterior to EC-VAE (ablation) | EC-VAE · CIFAR-10 | FID: 5.20 -> 5.81 | Hurt |
| add calibrated posterior to EC-VAE (ablation) | EC-VAE · CIFAR-10 | ELBO: 652.59 -> 1072.40 | Improved |
| add EMA to EC-VAE (ablation) | EC-VAE · CIFAR-10 | FID: 8.02 -> 5.20 | Improved |
Closest papers by meaning
- Explicitly Minimizing the Blur Error of Variational Autoencoders (ICLR 2023 · similarity 0.88)
- Diagnosing and Enhancing VAE Models (ICLR 2019 · similarity 0.88)
- Simple and Effective VAE Training with Calibrated Decoders (ICML 2021 · similarity 0.88)
- Unscented Autoencoder (ICML 2023 · similarity 0.88)
- A Contrastive Learning Approach for Training Variational Autoencoder Priors (NeurIPS 2021 · similarity 0.88)
- EQ-VAE: Equivariance Regularized Latent Space for Improved Generative Image Modeling (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
